【PaddlePaddle Hackathon 10】Add Doc-QnA demo with PaddleOCR-VL + OpenVINO

[bob798]

Summary

Add a new Document Q&A demo that combines PaddleOCR-VL with OpenVINO for intelligent document understanding and RAG-based question answering.

• PaddleOCR-VL (OpenVINO): Parses scanned PDFs, tables, and formulas into structured Markdown
• Table-aware chunking: Each table row carries its header for precise cell-level lookup
• Qwen3-Embedding-0.6B-int8 (OpenVINO): 1024-dim multilingual vector indexing via ChromaDB
• Qwen3-1.7B-int4 (OpenVINO GenAI): Answer generation with [doc_name p.page] source citations
• All-OpenVINO inference: OCR, Embedding, and LLM — no PyTorch at runtime
• CPU-friendly: ~3.5s/question on Intel i5

Demo output example

Q: A100 型号的工作温度范围是多少？
A: A100 型号的工作温度范围是 -20~70℃ [spec_with_tables p.1]
⏱  embed=75ms  retrieve=1.4ms  llm=3397ms  total=3474ms  tps=10.3

One-click run

cd demos/doc_qna_demo
pip install -r requirements.txt
python main.py

Models are auto-downloaded from HuggingFace on first run (~1.6 GB total for Embedding + LLM).

Files added

demos/doc_qna_demo/
├── main.py                    # Single entry point
├── README.md                  # Setup + usage + architecture
├── requirements.txt           # Pinned dependencies
├── setup/
│   ├── install.bat            # Windows one-click install + run
│   └── install.sh             # Linux/macOS one-click install + run
├── src/                       # Core modules
│   ├── embedding.py           # Qwen3-Embedding OpenVINO wrapper
│   ├── llm.py                 # Qwen3-1.7B OpenVINO GenAI wrapper
│   ├── rag.py                 # RAG orchestration
│   ├── vector_store.py        # ChromaDB persistence
│   ├── chunker.py             # Table-aware chunking
│   ├── doc_parser.py          # PaddleOCR-VL document parsing
│   ├── pdf_preprocessor.py    # PDF text-layer detection
│   ├── inference.py           # OpenVINO inference wrapper
│   └── pipeline.py            # Phase 2 end-to-end pipeline
├── data/
│   └── demo_questions.txt     # 5 business questions for demo
└── results/phase2/            # Pre-generated chunks (skip OCR step)

Test plan

• Verified on Windows (Python 3.12, Intel i5 CPU)
• pip install -r requirements.txt + python main.py runs end-to-end
• 5/5 questions answered with correct citations
• Performance: 3.5s/question avg on CPU
• Linux/macOS verification (pending)
• GPU device test (pending)